Neuro-architecture-motivated ANNs and cortical parcellation

An overview by Matsumoto et al. (1999) highlights “the shifting emphasis in the modelling of neural systems towards more neuroarchitecture-motivated systems”. There is a widespread assumption that the brain has evolved with a highly organised modular architecture covering a variety of anatomical scales reflecting similar variation in functional levels of neural processing. Currently, there exist no general methods for setting suitable structural constraints for neural networks which are dedicated to specific tasks. Modular constraints on architecture and connectivity, as found in the brain, and interpreted within a definite behavioural context, may provide useful guidelines for the effective design of artificial neural systems which have the capability to deal with real-world problems. In a search for such guidelines our work focuses on the ontogeny and phylogeny of parcellation of the cerebral cortex. Emphasis is placed on processes underlying the emergence of parcellation on an evolutionary time scale. Slow learning and accelerated evolution, involving a form of inheritance of acquired characteristics, are assigned critical roles. The dominant neuro-anatomical view is that the majority of cerebral cortical tissue is largely equipotent early in epigenesis. Modularity of function is, accordingly, a result of intragenerational experience. With Johnson (1997), we adopt a more probabilistic epigenetic view in which certain cortical areas have a detailed architecture slightly different from the basic neural structure common to the cortex. This makes them the most efficient at processing certain types of input and increases the probability of the appearance of appropriate functional modularity with experience. Our theory includes processes sufficient to account for the development of functionally tilted cortical area architectures in the course of evolution